/* spl_syscall.c - built-in syscall implementations + registration
 *
 * All syscalls validate nargs, cast spl_val_t args to the expected C types,
 * execute, and return the result as spl_val_t.
 *
 * Simple syscalls use the SYSCALL_N macros; complex ones are written
 * manually but follow the same template.
 */

#include "spl_syscall.h"
#include "include/core_map.h"
#include "include/core_vec.h"
#include "spl_mcode.h"
#include "spl_vm.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>

/* =================================================================
 * Macro templates
 *
 * Each SYSCALL_N macro:
 *  1. Validates nargs (prints error and returns -1 on mismatch)
 *  2. Casts args[n] from spl_val_t to the specified C type via
 *     (type)(uintptr_t) - this handles both integer and pointer types
 *  3. Evaluates the expression and returns the result as spl_val_t
 * ================================================================= */

#define CHECK_NARGS(fname, expected)                                                               \
    do {                                                                                           \
        if (nargs != (expected)) {                                                                 \
            fprintf(stderr, fname ": expected " #expected " args, got %d\n", nargs);               \
            return -1;                                                                             \
        }                                                                                          \
    } while (0)

#define SYSCALL_0(name, ret_expr)                                                                  \
    static spl_vm_val_t name(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {                       \
        (void)_vm;                                                                                 \
        CHECK_NARGS(#name, 0);                                                                     \
        (void)args;                                                                                \
        return (spl_vm_val_t)(uintptr_t)(ret_expr);                                                \
    }

#define SYSCALL_1(name, t1, ret_expr)                                                              \
    static spl_vm_val_t name(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {                       \
        (void)_vm;                                                                                 \
        CHECK_NARGS(#name, 1);                                                                     \
        t1 a1 = (t1)(uintptr_t)args[0];                                                            \
        return (spl_vm_val_t)(uintptr_t)(ret_expr);                                                \
    }

#define SYSCALL_2(name, t1, t2, ret_expr)                                                          \
    static spl_vm_val_t name(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {                       \
        (void)_vm;                                                                                 \
        CHECK_NARGS(#name, 2);                                                                     \
        t1 a1 = (t1)(uintptr_t)args[0];                                                            \
        t2 a2 = (t2)(uintptr_t)args[1];                                                            \
        return (spl_vm_val_t)(uintptr_t)(ret_expr);                                                \
    }

#define SYSCALL_3(name, t1, t2, t3, ret_expr)                                                      \
    static spl_vm_val_t name(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {                       \
        (void)_vm;                                                                                 \
        CHECK_NARGS(#name, 3);                                                                     \
        t1 a1 = (t1)(uintptr_t)args[0];                                                            \
        t2 a2 = (t2)(uintptr_t)args[1];                                                            \
        t3 a3 = (t3)(uintptr_t)args[2];                                                            \
        return (spl_vm_val_t)(uintptr_t)(ret_expr);                                                \
    }

#define SYSCALL_4(name, t1, t2, t3, t4, ret_expr)                                                  \
    static spl_vm_val_t name(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {                       \
        (void)_vm;                                                                                 \
        CHECK_NARGS(#name, 4);                                                                     \
        t1 a1 = (t1)(uintptr_t)args[0];                                                            \
        t2 a2 = (t2)(uintptr_t)args[1];                                                            \
        t3 a3 = (t3)(uintptr_t)args[2];                                                            \
        t4 a4 = (t4)(uintptr_t)args[3];                                                            \
        return (spl_vm_val_t)(uintptr_t)(ret_expr);                                                \
    }

/* syscall / VM 内部统一地址检查宏。
 * 失败时打印 ctx + 地址 + size + ip，置 native_error，return 0。
 * 要求所在函数有可用的 _vm 变量。 */
#define ADDRCHECK(vm, ptr, size, ctx)                                                              \
    do {                                                                                           \
        uintptr_t _acp = (uintptr_t)(ptr);                                                         \
        usize _acs = (usize)(size);                                                                \
        if (!spl_vm_addr_valid(_acp, _acs)) {                                                      \
            fprintf(stderr, "%s: invalid address %p size=%zu (ip=%zu fp=%zu sp=%zu)\n", (ctx),     \
                    (void *)_acp, _acs, (vm)->ip, (vm)->fp, (vm)->sp);                             \
            spl_vm_backtrace((vm), (vm)->fp);                                                      \
            (vm)->exit_code = 2;                                                                   \
            return 0;                                                                              \
        }                                                                                          \
    } while (0)

/* syscall 内部：把当前栈和调用栈打印出来，供出错时定位 */
#define SYSCALL_DUMP(vm)                                                                           \
    do {                                                                                           \
        fprintf(stderr, "  --- syscall dump (sp=%zu fp=%zu cp=%zu ip=%zu) ---\n", (vm)->sp,        \
                (vm)->fp, (vm)->cp, (vm)->ip);                                                     \
        spl_vm_backtrace((vm), (vm)->fp);                                                          \
        spl_vm_stackdump((vm), (vm)->sp);                                                          \
    } while (0)

/* syscall 内部：报告致命错误，dump 后让 VM 停在当前 ip */
#define SYSCALL_FAIL(vm, ...)                                                                      \
    do {                                                                                           \
        fprintf(stderr, "syscall error: ");                                                        \
        fprintf(stderr, __VA_ARGS__);                                                              \
        fprintf(stderr, "\n");                                                                     \
        SYSCALL_DUMP(vm);                                                                          \
        (vm)->exit_code = 1;                                                                       \
        return 0;                                                                                  \
    } while (0)

/* =================================================================
 * OS syscalls
 * ================================================================= */

/* vm_exit - terminate process with the given exit code */
SYSCALL_1(vm_exit, int, (exit(a1), (spl_vm_val_t)0))

/* vm_putchar - write a single character to stdout */
SYSCALL_1(vm_putchar, int, putchar(a1))

/* vm_getchar - read a single character from stdin */
SYSCALL_0(vm_getchar, getchar())

/* vm_putint - print an integer to stdout */
SYSCALL_1(vm_putint, int, (fprintf(stdout, "%d", a1), (spl_vm_val_t)0))

/* vm_putstr - print a string to stdout (no trailing newline) */
SYSCALL_1(vm_putstr, const char *, (fputs(a1, stdout), (spl_vm_val_t)0))

/* vm_fopen - open a file, returns FILE* as spl_val_t */
SYSCALL_2(vm_fopen, const char *, const char *, (uintptr_t)fopen(a1, a2))

/* vm_fclose - close a file, returns 0 on success */
SYSCALL_1(vm_fclose, FILE *, fclose(a1))

/* vm_fread - read from file, returns number of items read */
SYSCALL_4(vm_fread, void *, size_t, size_t, FILE *, fread(a1, a2, a3, a4))

/* vm_fwrite - write to file, returns number of items written */
SYSCALL_4(vm_fwrite, const void *, size_t, size_t, FILE *, fwrite(a1, a2, a3, a4))

/* vm_fsize - get file size in bytes */
static spl_vm_val_t vm_fsize(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_fsize", 1);
    FILE *f = (FILE *)(uintptr_t)args[0];
    long cur = ftell(f);
    fseek(f, 0, SEEK_END);
    long sz = ftell(f);
    fseek(f, cur, SEEK_SET);
    return (spl_vm_val_t)(uintptr_t)sz;
}

SYSCALL_0(vm_stdin, stdin)
SYSCALL_0(vm_stdout, stdout)
SYSCALL_0(vm_stderr, stderr)

/* vm_read_file - read entire file into a malloc'd, null-terminated buffer */
static spl_vm_val_t vm_read_file(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_read_file", 1);
    const char *path = (const char *)(uintptr_t)args[0];
    if (path == nullptr) {
        fprintf(stderr, "filepath can't be null");
        return 0;
    }
    FILE *f = fopen(path, "rb");
    if (!f) {
        fprintf(stderr, "filepath %s can't be open", path);
        return 1;
    }
    fseek(f, 0, SEEK_END);
    long sz = ftell(f);
    fseek(f, 0, SEEK_SET);
    char *buf = (char *)malloc((size_t)sz + 1);
    if (!buf) {
        fclose(f);
        return 0;
    }
    size_t nread = fread(buf, 1, (size_t)sz, f);
    fclose(f);
    buf[nread] = '\0';
    return (spl_vm_val_t)(uintptr_t)buf;
}

static spl_vm_val_t vm_alloc(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_alloc", 1);
    size_t sz = (size_t)args[0];
    if (sz == 0)
        SYSCALL_FAIL(_vm, "vm_alloc: zero size");
    if (sz > (size_t)1 << 32)
        SYSCALL_FAIL(_vm, "vm_alloc: size too large (%zu)", sz);
    void *p = malloc(sz);
    if (!p)
        SYSCALL_FAIL(_vm, "vm_alloc: malloc(%zu) failed", sz);
    return (spl_vm_val_t)(uintptr_t)p;
}

static spl_vm_val_t vm_free(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_free", 1);
    void *p = (void *)(uintptr_t)args[0];
    if (p)
        free(p);
    return 0;
}

static spl_vm_val_t vm_realloc(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_realloc", 2);
    void *p = (void *)(uintptr_t)args[0];
    size_t sz = (size_t)args[1];
    if (!p)
        SYSCALL_FAIL(_vm, "vm_realloc: NULL ptr");
    if (sz == 0)
        SYSCALL_FAIL(_vm, "vm_realloc: zero size");
    ADDRCHECK(_vm, p, 1, "vm_realloc");
    void *np = realloc(p, sz);
    if (!np)
        SYSCALL_FAIL(_vm, "vm_realloc: realloc(%p, %zu) failed", p, sz);
    return (spl_vm_val_t)(uintptr_t)np;
}

static spl_vm_val_t vm_memcpy(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_memcpy", 3);
    void *dst = (void *)(uintptr_t)args[0];
    const void *src = (const void *)(uintptr_t)args[1];
    size_t size = (size_t)args[2];
    if (size == 0)
        return (spl_vm_val_t)(uintptr_t)dst;
    ADDRCHECK(_vm, dst, size, "vm_memcpy(dst)");
    ADDRCHECK(_vm, src, size, "vm_memcpy(src)");
    memcpy(dst, src, size);
    return (spl_vm_val_t)(uintptr_t)dst;
}

static spl_vm_val_t vm_memcmp(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_memcmp", 3);
    const void *a = (const void *)(uintptr_t)args[0];
    const void *b = (const void *)(uintptr_t)args[1];
    size_t size = (size_t)args[2];
    if (size == 0)
        return 0;
    ADDRCHECK(_vm, a, size, "vm_memcmp(a)");
    ADDRCHECK(_vm, b, size, "vm_memcmp(b)");
    return (spl_vm_val_t)memcmp(a, b, size);
}

static spl_vm_val_t vm_printf(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    if (nargs <= 0)
        return 0;

    /* args[0] 是 SPL 的 []u8 format 参数。VM 里 slice 作为参数传递时，
       栈上压的是"指向 slice 结构体的指针"，结构为 {ptr:*u8, len:u64} */
    const spl_vm_val_t *fmt_slice = (const spl_vm_val_t *)args[0];
    const char *fmt = (const char *)fmt_slice[0];
    usize fmt_len = (usize)fmt_slice[1];

    if (nargs <= 1) {
        fwrite(fmt, 1, fmt_len, stdout);
        return (spl_vm_val_t)nargs;
    }

    typedef VEC(char) string_builder_t;
    string_builder_t buffer;
    vec_init(buffer);

    int arg_idx = 1; /* 0 是 format */
    char tmp_buf[32];

    for (usize i = 0; i < fmt_len; ++i) {
        if (fmt[i] != '%') {
            vec_push(buffer, fmt[i]);
            continue;
        }
        if (++i >= fmt_len)
            break;
        if (arg_idx >= nargs)
            break;

        switch (fmt[i]) {
        case 'd': {
            memset(tmp_buf, 0, sizeof(tmp_buf));
            snprintf(tmp_buf, sizeof(tmp_buf), "%zd", (isize)args[arg_idx]);
            for (usize j = 0; tmp_buf[j] != '\0'; ++j)
                vec_push(buffer, tmp_buf[j]);
            arg_idx++;
            break;
        }
        case 'c':
            vec_push(buffer, (char)args[arg_idx]);
            arg_idx++;
            break;
        case 's': {
            const spl_vm_val_t *s = (const spl_vm_val_t *)args[arg_idx];
            const char *p = (const char *)s[0];
            usize l = (usize)s[1];
            for (usize j = 0; j < l; ++j)
                vec_push(buffer, p[j]);
            arg_idx++;
            break;
        }
        default:
            vec_push(buffer, '%');
            vec_push(buffer, fmt[i]);
            break;
        }
    }

    if (vec_size(buffer) > 0)
        fwrite(buffer.data, 1, vec_size(buffer), stdout);
    vec_free(buffer);
    return (spl_vm_val_t)nargs;
}

/* 探测接口（syscall 风格，非必选，逐步扩展）。
 * 不假设一定可用：不支持时统一返回 0。
 *   case 0 -> 是否存在 time：1 = 存在（则 case 1/2 有效），0 = 不存在
 *   case 1 -> 当前秒（time_t）
 *   case 2 -> 当前纳秒部分（tv_nsec）
 *   case 3 -> VM 已执行指令数（参考量；解码/执行恒定时可换算时间）
 *   case 4 -> 完整纳秒时间戳（单次调用，避免 sec/ns 跨秒竞态；
 *             纪元纳秒约 1.8e18 < u64 上限，usize 足够，无需 i128）
 * 新增探测点先加 case，再在 SPL 侧逐步验证。 */
static spl_vm_val_t vm_timestamp(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_timestamp", 1);
    spl_vm_val_t arg = args[0];
    struct timespec ts;
    int has_time = (timespec_get(&ts, TIME_UTC) == TIME_UTC);
    switch (arg) {
    case 0:
        return has_time ? 1 : 0;
    case 1:
        return has_time ? (spl_vm_val_t)ts.tv_sec : 0;
    case 2:
        return has_time ? (spl_vm_val_t)ts.tv_nsec : 0;
    case 3:
        if (!_vm)
            return 0;
        return (spl_vm_val_t)_vm->steps;
    case 5:
        if (!_vm)
            return 0;
        return (spl_vm_val_t)_vm->ncalls;
    case 4:
        return has_time ? (spl_vm_val_t)((unsigned long long)ts.tv_sec * 1000000000ull +
                                         (unsigned long long)ts.tv_nsec)
                        : 0;
    default:
        fprintf(stderr, "vm_timestamp: invalid arg %zu\n", (size_t)arg);
        return 0;
    }
}

/* =================================================================
 * VM syscalls (for comptime - compile-time code execution)
 *
 * These let SPL code create isolated sub-VMs, load compiled .sir
 * programs into them, push arguments, call functions, and run.
 * ================================================================= */

/* vm_new - create a new sub-VM instance, returns spl_vm_t* */
static spl_vm_val_t vm_new(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_new", 0);
    (void)args;
    spl_vm_t *vm = (spl_vm_t *)malloc(sizeof(spl_vm_t));
    if (!vm)
        return 0;
    spl_vm_init(vm);
    return (spl_vm_val_t)(uintptr_t)vm;
}

/* vm_drop - destroy a sub-VM and its loaded program */
static spl_vm_val_t vm_drop(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_drop", 1);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    if (!vm)
        return 0;
    if (vm->prog) {
        spl_prog_drop(vm->prog);
        free(vm->prog);
        vm->prog = NULL;
    }
    spl_vm_drop(vm);
    free(vm);
    return 0;
}

/* vm_load - load a .sir file, register syscalls, return spl_prog_t* */
static spl_vm_val_t vm_load(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_load", 2);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    const char *path = (const char *)(uintptr_t)args[1];
    if (!vm || !path)
        return -1;

    spl_prog_t *prog = (spl_prog_t *)malloc(sizeof(spl_prog_t));
    if (!prog)
        return -1;
    if (spl_prog_load_from_file(path, prog) != 0) {
        free(prog);
        return -1;
    }
    spl_syscall_register(prog);
    if (spl_vm_load_prog(vm, prog) != 0) {
        spl_prog_drop(prog);
        free(prog);
        return -1;
    }
    return (spl_vm_val_t)(uintptr_t)prog;
}

/* vm_load_from_buffer - 从内存加载字节码（comptime 内嵌 VM 用）。
 * 参数 1 是 slice 结构 { ptr, len } 的地址（聚合参数按指针传递）。 */
static spl_vm_val_t vm_load_from_buffer(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_load_from_buffer", 3);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    const char *buf = (const char *)(uintptr_t)args[1];
    usize len = (usize)args[2];
    spl_prog_t *prog;
    if (!vm || !buf)
        return -1;
    prog = (spl_prog_t *)malloc(sizeof(spl_prog_t));
    if (!prog)
        return -1;
    if (spl_prog_load_from_buffer(buf, len, prog) != 0) {
        free(prog);
        return -1;
    }
    spl_syscall_register(prog);
    if (spl_vm_load_prog(vm, prog) != 0) {
        spl_prog_drop(prog);
        free(prog);
        return -1;
    }
    return 0;
}

/* vm_push - push a value onto the sub-VM's stack (for passing arguments) */
static spl_vm_val_t vm_push(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_push", 2);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    spl_vm_val_t val = args[1];
    if (!vm)
        return -1;
    if (vm->sp >= vm->config.max_stack_depth) {
        fprintf(stderr, "vm_push: stack overflow\n");
        return -1;
    }
    vm->stacks.data[vm->sp++] = val;
    return 0;
}

/* vm_call - call a function by name with pre-pushed arguments
 *
 * Args (3): vm, function_name, nargs
 * The arguments should already be on the sub-VM's stack via vm_push.
 *
 * Mirrors the CALL instruction's frame setup:
 *   - Pushes a sentinel frame (saved_ip=-1) if cp==0 so RET halts properly
 *   - Pushes the actual call frame
 *   - Sets fp = sp - nargs  (so arg0 = data[fp], arg1 = data[fp+1], ...)
 *   - Sets ip to the function address
 */
static spl_vm_val_t vm_call(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_call", 3);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    const char *name = (const char *)(uintptr_t)args[1];
    spl_vm_val_t call_nargs = args[2];
    if (!vm || !name)
        return -1;

    spl_vm_func_t *func = spl_prog_get_func(vm->prog, name);
    if (!func) {
        fprintf(stderr, "vm_call: function '%s' not found\n", name);
        return -1;
    }
    if (vm->cp >= vm->config.max_call_depth) {
        fprintf(stderr, "vm_call: call stack overflow\n");
        return -1;
    }

    /* 入口调用：ip 置 -1 作为 sentinel，RET 时 halt 并把返回值写入 exit_code。
     * 实参须已由 vm_push 压栈（sp = nargs）。 */
    vm->ip = (uintptr_t)-1;
    vm->cp = 0;
    vm->fp = 0;
    vm->frames.data[0].saved_sp = vm->sp - call_nargs;
    vm->frames.data[0].saved_fp = vm->fp;
    vm->frames.data[0].saved_ip = vm->ip;
    vm->frames.data[0].nargs = call_nargs;
    vm->cp = 1;
    vm->ip = (uintptr_t)func->address;
    vm->fp = vm->sp - call_nargs;
    return 0;
}

/* vm_run - run a sub-VM to completion, returns exit_code（入口 RET 的值） */
static spl_vm_val_t vm_run(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_run", 1);
    spl_vm_t *vm = (spl_vm_t *)(uintptr_t)args[0];
    if (!vm)
        return -1;
    spl_vm_run_until(vm, 0);
    return (spl_vm_val_t)vm->exit_code;
}

/* =================================================================
 * Registration
 *
 * Called after loading a .sir file.  Matches native declarations in
 * prog->natives against known syscall names and hooks up impl_fn.
 * ================================================================= */

/* =================================================================
 * Native std containers (spike: Vec valid/push_raw/get)
 *
 * 布局镜像 stage2/utils/vec.spl 的 Vec:
 *   raw, data, len, cap, elem_size, elem_align  (均为机器字)
 * 结构体按值传递/复制由 SPL 负责，native 只按指针操作。
 * ================================================================= */

typedef struct {
    uintptr_t raw;
    uintptr_t data;
    uintptr_t len;
    uintptr_t cap;
    uintptr_t elem_size;
    uintptr_t elem_align;
} spl_vec_t;

static int spl_vec_is_valid(const spl_vec_t *v) {
    if (v->elem_size == 0)
        return 0;
    uintptr_t a = v->elem_align;
    if (a == 0 || (a & (a - 1)) != 0)
        return 0;
    if (v->len > v->cap)
        return 0;
    if (v->cap == 0)
        return v->raw == 0 && v->data == 0 && v->len == 0;
    if (v->raw == 0 || v->data == 0)
        return 0;
    return (v->data & (a - 1)) == 0;
}

static spl_vm_val_t vm_vec_valid(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_valid", 1);
    const spl_vec_t *v = (const spl_vec_t *)(uintptr_t)args[0];
    return (spl_vm_val_t)(v ? spl_vec_is_valid(v) : 0);
}

static spl_vm_val_t vm_vec_push_raw(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_push_raw", 2);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    const void *elem = (const void *)(uintptr_t)args[1];
    if (!v || !elem)
        return 0;
    if (v->len < v->cap) {
        memcpy((void *)(v->data + v->len * v->elem_size), elem, v->elem_size);
        v->len++;
        return 1;
    }
    return 0;
}

/* out 接收元素指针；返回 0 成功，1 无效参数，2 越界。 */
static int spl_vec_reserve_raw(spl_vec_t *v, uintptr_t need);

static spl_vm_val_t vm_vec_push(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_push", 2);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    const void *elem = (const void *)(uintptr_t)args[1];
    if (!v || !elem || !spl_vec_is_valid(v))
        return 1;
    if (v->len < v->cap) {
        memcpy((void *)(v->data + v->len * v->elem_size), elem, v->elem_size);
        v->len++;
        return 0;
    }
    int rc = spl_vec_reserve_raw(v, v->len + 1);
    if (rc)
        return rc;
    memmove((void *)(v->data + v->len * v->elem_size), elem, v->elem_size);
    v->len++;
    return 0;
}

static spl_vm_val_t vm_vec_get(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_get", 3);
    const spl_vec_t *v = (const spl_vec_t *)(uintptr_t)args[0];
    uintptr_t index = (uintptr_t)args[1];
    uintptr_t *out = (uintptr_t *)(uintptr_t)args[2];
    if (!v || !out)
        return 1;
    if (!spl_vec_is_valid(v))
        return 1;
    if (index >= v->len)
        return 2;
    *out = v->data + index * v->elem_size;
    return 0;
}

/* 错误码统一：0 ok / 1 InvalidArgument / 2 OutOfRange / 3 Overflow / 4 InvalidAlignment */

static int spl_align_ok(uintptr_t a) { return a != 0 && (a & (a - 1)) == 0; }

static int spl_grow_capacity(uintptr_t current, uintptr_t min_cap, uintptr_t need,
                            uintptr_t *out) {
    uintptr_t cap = current < min_cap ? min_cap : current;
    uintptr_t limit = ~(uintptr_t)0;
    while (cap < need) {
        if (cap > limit / 2) {
            cap = need;
            break;
        }
        cap = cap * 2;
    }
    *out = cap;
    return 0;
}

static int spl_vec_allocate(spl_vec_t *out, uintptr_t capacity, uintptr_t elem_size,
                            uintptr_t elem_align) {
    if (elem_size == 0)
        return 1;
    if (!spl_align_ok(elem_align))
        return 4;
    out->raw = 0;
    out->data = 0;
    out->len = 0;
    out->cap = 0;
    out->elem_size = elem_size;
    out->elem_align = elem_align;
    if (capacity == 0)
        return 0;
    if (capacity > ~(uintptr_t)0 / elem_size)
        return 3;
    uintptr_t bytes = capacity * elem_size;
    uintptr_t mask = elem_align - 1;
    if (bytes > ~(uintptr_t)0 - mask)
        return 3;
    uintptr_t total = bytes + mask;
    void *raw = malloc(total);
    if (!raw)
        return 3;
    uintptr_t addr = (uintptr_t)raw;
    if (addr > ~(uintptr_t)0 - mask) {
        free(raw);
        return 3;
    }
    out->raw = addr;
    out->data = (addr + mask) & ~mask;
    out->cap = capacity;
    return 0;
}

static spl_vm_val_t vm_vec_allocate(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_allocate", 4);
    spl_vec_t *out = (spl_vec_t *)(uintptr_t)args[0];
    if (!out)
        return 1;
    return (spl_vm_val_t)spl_vec_allocate(out, (uintptr_t)args[1], (uintptr_t)args[2],
                                          (uintptr_t)args[3]);
}

static int spl_vec_reserve_raw(spl_vec_t *v, uintptr_t need) {
    if (need <= v->cap)
        return 0;
    uintptr_t new_cap = 0;
    spl_grow_capacity(v->cap, 8, need, &new_cap);
    spl_vec_t fresh;
    int rc = spl_vec_allocate(&fresh, new_cap, v->elem_size, v->elem_align);
    if (rc)
        return rc;
    uintptr_t used = 0;
    if (v->len != 0) {
        if (v->len > ~(uintptr_t)0 / v->elem_size) {
            free((void *)fresh.raw);
            return 3;
        }
        used = v->len * v->elem_size;
    }
    if (used != 0)
        memcpy((void *)fresh.data, (void *)v->data, used);
    fresh.len = v->len;
    uintptr_t old = v->raw;
    *v = fresh;
    if (old)
        free((void *)old);
    return 0;
}

static spl_vm_val_t vm_vec_reserve_raw(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_reserve_raw", 2);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    if (!v)
        return 1;
    return (spl_vm_val_t)spl_vec_reserve_raw(v, (uintptr_t)args[1]);
}

static spl_vm_val_t vm_vec_append(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_append", 3);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    const void *elems = (const void *)(uintptr_t)args[1];
    uintptr_t count = (uintptr_t)args[2];
    if (!v || !spl_vec_is_valid(v))
        return 1;
    if (count == 0)
        return 0;
    if (!elems)
        return 1;
    if (count > ~(uintptr_t)0 / v->elem_size)
        return 3;
    uintptr_t copy_bytes = count * v->elem_size;
    if (v->len > ~(uintptr_t)0 - count)
        return 3;
    uintptr_t new_len = v->len + count;
    uintptr_t used = v->len * v->elem_size;
    int source_inside = 0;
    uintptr_t source_offset = 0;
    if (used != 0) {
        uintptr_t base = v->data;
        uintptr_t source = (uintptr_t)elems;
        if (source >= base && source - base < used) {
            source_offset = source - base;
            if (copy_bytes > used - source_offset)
                return 2;
            source_inside = 1;
        }
    }
    int rc = spl_vec_reserve_raw(v, new_len);
    if (rc)
        return rc;
    if (source_inside)
        elems = (const void *)(v->data + source_offset);
    memmove((void *)(v->data + used), elems, copy_bytes);
    v->len = new_len;
    return 0;
}

static spl_vm_val_t vm_vec_set(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_set", 3);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    uintptr_t index = (uintptr_t)args[1];
    const void *elem = (const void *)(uintptr_t)args[2];
    if (!v || !elem || !spl_vec_is_valid(v))
        return 1;
    if (index >= v->len)
        return 2;
    memmove((void *)(v->data + index * v->elem_size), elem, v->elem_size);
    return 0;
}

static spl_vm_val_t vm_vec_pop(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_pop", 2);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    void *out = (void *)(uintptr_t)args[1];
    if (!v || !out || !spl_vec_is_valid(v))
        return 1;
    if (v->len == 0)
        return 2;
    memmove(out, (void *)(v->data + (v->len - 1) * v->elem_size), v->elem_size);
    v->len--;
    return 0;
}

static spl_vm_val_t vm_vec_clone(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_clone", 2);
    spl_vec_t *out = (spl_vec_t *)(uintptr_t)args[0];
    const spl_vec_t *src = (const spl_vec_t *)(uintptr_t)args[1];
    if (!out || !src || !spl_vec_is_valid(out) || !spl_vec_is_valid(src))
        return 1;
    if (out->elem_size != src->elem_size || out->elem_align != src->elem_align)
        return 1;
    spl_vec_t fresh;
    fresh.raw = 0;
    fresh.data = 0;
    fresh.len = 0;
    fresh.cap = 0;
    fresh.elem_size = src->elem_size;
    fresh.elem_align = src->elem_align;
    int rc = spl_vec_reserve_raw(&fresh, src->len);
    if (rc)
        return rc;
    uintptr_t bytes = src->len * src->elem_size;
    if (bytes != 0)
        memcpy((void *)fresh.data, (void *)src->data, bytes);
    fresh.len = src->len;
    if (out->raw)
        free((void *)out->raw);
    *out = fresh;
    return 0;
}

static spl_vm_val_t vm_vec_drop(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_vec_drop", 1);
    spl_vec_t *v = (spl_vec_t *)(uintptr_t)args[0];
    if (!v)
        return 0;
    uintptr_t es = v->elem_size;
    uintptr_t ea = v->elem_align;
    if (v->raw)
        free((void *)v->raw);
    v->raw = 0;
    v->data = 0;
    v->len = 0;
    v->cap = 0;
    v->elem_size = es;
    v->elem_align = ea;
    return 0;
}

static spl_vm_val_t vm_memmove(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    CHECK_NARGS("vm_memmove", 3);
    void *dst = (void *)(uintptr_t)args[0];
    const void *src = (const void *)(uintptr_t)args[1];
    size_t size = (size_t)args[2];
    if (size == 0)
        return (spl_vm_val_t)(uintptr_t)dst;
    ADDRCHECK(_vm, dst, size, "vm_memmove(dst)");
    ADDRCHECK(_vm, src, size, "vm_memmove(src)");
    memmove(dst, src, size);
    return (spl_vm_val_t)(uintptr_t)dst;
}

/* =================================================================
 * Native Map（开放寻址 + 墓碑），布局镜像 stage2/utils/map.spl 的 Map。
 * hash_kind: 0 = usize；1 = ByteView（按内容）；2 = 常量（仅测试用）。
 * ================================================================= */

typedef struct {
    spl_vec_t keys;
    spl_vec_t vals;
    spl_vec_t states;
    uintptr_t size;
    uintptr_t key_size;
    uintptr_t val_size;
    uintptr_t hash_kind;
} spl_map_t;

static void spl_vec_release(spl_vec_t *v) {
    uintptr_t es = v->elem_size;
    uintptr_t ea = v->elem_align;
    if (v->raw)
        free((void *)v->raw);
    v->raw = 0;
    v->data = 0;
    v->len = 0;
    v->cap = 0;
    v->elem_size = es;
    v->elem_align = ea;
}

static uintptr_t spl_hash_usize_value(uintptr_t value) {
    uintptr_t prime = 1099511628211ull;
    uintptr_t h = value;
    h = (h ^ (h >> 30)) * prime;
    h = h ^ (h >> 27);
    h = h * prime;
    h = h ^ (h >> 31);
    if (h == 0)
        return prime;
    return h;
}

static uintptr_t spl_hash_bytes(const void *data, uintptr_t len) {
    const unsigned char *p = (const unsigned char *)data;
    uintptr_t h = 14695981039346656037ull;
    uintptr_t prime = 1099511628211ull;
    for (uintptr_t i = 0; i < len; i++)
        h = (h ^ (uintptr_t)p[i]) * prime;
    return h;
}

static uintptr_t spl_map_hash(const spl_map_t *m, const void *key) {
    if (m->hash_kind == 0)
        return spl_hash_usize_value(*(const uintptr_t *)key);
    if (m->hash_kind == 2)
        return 3;
    const uintptr_t *v = (const uintptr_t *)key; /* ByteView { ptr, len } */
    return spl_hash_bytes((const void *)v[0], v[1]);
}

static int spl_map_eq(const spl_map_t *m, const void *a, const void *b) {
    if (m->hash_kind == 0 || m->hash_kind == 2)
        return *(const uintptr_t *)a == *(const uintptr_t *)b;
    const uintptr_t *x = (const uintptr_t *)a;
    const uintptr_t *y = (const uintptr_t *)b;
    if (x[1] != y[1])
        return 0;
    if (x[1] == 0)
        return 1;
    return memcmp((const void *)x[0], (const void *)y[0], x[1]) == 0;
}

static int spl_map_valid(const spl_map_t *m) {
    if (m->key_size == 0 || m->val_size == 0)
        return 0;
    if (m->keys.len != m->states.len || m->vals.len != m->states.len)
        return 0;
    if (m->size > m->states.len)
        return 0;
    return 1;
}

static uintptr_t spl_map_find_slot(const spl_map_t *m, const void *key) {
    if (m->states.cap == 0)
        return ~(uintptr_t)0;
    uintptr_t mask = m->states.cap - 1;
    uintptr_t i = spl_map_hash(m, key) & mask;
    unsigned char *st = (unsigned char *)m->states.data;
    uintptr_t probes = 0;
    while (probes < m->states.cap && st[i] != 0) {
        if (st[i] == 1 && spl_map_eq(m, (const void *)(m->keys.data + i * m->key_size), key))
            return i;
        i = (i + 1) & mask;
        probes++;
    }
    return ~(uintptr_t)0;
}

static void spl_map_place(spl_map_t *m, const void *key, const void *val) {
    uintptr_t mask = m->states.cap - 1;
    uintptr_t i = spl_map_hash(m, key) & mask;
    unsigned char *st = (unsigned char *)m->states.data;
    uintptr_t first_deleted = ~(uintptr_t)0;
    uintptr_t probes = 0;
    while (probes < m->states.cap && st[i] != 0) {
        if (st[i] == 2 && first_deleted == ~(uintptr_t)0)
            first_deleted = i;
        i = (i + 1) & mask;
        probes++;
    }
    if (first_deleted != ~(uintptr_t)0)
        i = first_deleted;
    memcpy((void *)(m->keys.data + i * m->key_size), key, m->key_size);
    memcpy((void *)(m->vals.data + i * m->val_size), val, m->val_size);
    st[i] = 1;
    m->size++;
}

static int spl_map_reserve(spl_map_t *m, uintptr_t need) {
    if (!spl_map_valid(m))
        return 1;
    if (need <= m->states.cap)
        return 0;
    uintptr_t new_cap = 0;
    spl_grow_capacity(m->states.cap, 8, need, &new_cap);

    spl_map_t fresh = *m;
    fresh.keys.raw = 0;
    fresh.keys.data = 0;
    fresh.keys.len = 0;
    fresh.keys.cap = 0;
    fresh.vals.raw = 0;
    fresh.vals.data = 0;
    fresh.vals.len = 0;
    fresh.vals.cap = 0;
    fresh.states.raw = 0;
    fresh.states.data = 0;
    fresh.states.len = 0;
    fresh.states.cap = 0;
    fresh.size = 0;

    int rc = spl_vec_reserve_raw(&fresh.keys, new_cap);
    if (rc)
        goto fail;
    rc = spl_vec_reserve_raw(&fresh.vals, new_cap);
    if (rc)
        goto fail;
    rc = spl_vec_reserve_raw(&fresh.states, new_cap);
    if (rc)
        goto fail;
    fresh.keys.len = new_cap;
    fresh.vals.len = new_cap;
    fresh.states.len = new_cap;
    memset((void *)fresh.states.data, 0, new_cap);

    for (uintptr_t i = 0; i < m->states.len; i++) {
        if (((unsigned char *)m->states.data)[i] == 1)
            spl_map_place(&fresh, (const void *)(m->keys.data + i * m->key_size),
                          (const void *)(m->vals.data + i * m->val_size));
    }

    spl_vec_release(&m->keys);
    spl_vec_release(&m->vals);
    spl_vec_release(&m->states);
    *m = fresh;
    return 0;
fail:
    spl_vec_release(&fresh.keys);
    spl_vec_release(&fresh.vals);
    spl_vec_release(&fresh.states);
    return rc;
}

static spl_vm_val_t vm_bytes_eq(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_bytes_eq", 4);
    const void *a = (const void *)(uintptr_t)args[0];
    uintptr_t alen = (uintptr_t)args[1];
    const void *b = (const void *)(uintptr_t)args[2];
    uintptr_t blen = (uintptr_t)args[3];
    if (alen != blen)
        return 0;
    if (alen == 0)
        return 1;
    if (!a || !b)
        return 0;
    return memcmp(a, b, alen) == 0;
}

static spl_vm_val_t vm_map_valid(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_valid", 1);
    const spl_map_t *m = (const spl_map_t *)(uintptr_t)args[0];
    return (spl_vm_val_t)(m ? spl_map_valid(m) : 0);
}

static spl_vm_val_t vm_map_reserve(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_reserve", 2);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    if (!m)
        return 1;
    return (spl_vm_val_t)spl_map_reserve(m, (uintptr_t)args[1]);
}

static spl_vm_val_t vm_map_put(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_put", 3);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    const void *key = (const void *)(uintptr_t)args[1];
    const void *val = (const void *)(uintptr_t)args[2];
    if (!m || !key || !val || !spl_map_valid(m))
        return 1;
    uintptr_t existing = spl_map_find_slot(m, key);
    if (existing != ~(uintptr_t)0) {
        memcpy((void *)(m->vals.data + existing * m->val_size), val, m->val_size);
        return 0;
    }
    if (m->states.cap == 0) {
        int rc = spl_map_reserve(m, 8);
        if (rc)
            return rc;
    } else {
        if (m->size > ~(uintptr_t)0 - 1)
            return 3;
        uintptr_t next_size = m->size + 1;
        if (next_size > ~(uintptr_t)0 / 128)
            return 3;
        uintptr_t scaled = next_size * 128;
        if (scaled / m->states.cap >= 70) {
            if (m->states.cap > ~(uintptr_t)0 / 2)
                return 3;
            int rc = spl_map_reserve(m, m->states.cap * 2);
            if (rc)
                return rc;
        }
    }
    spl_map_place(m, key, val);
    return 0;
}

static spl_vm_val_t vm_map_get(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_get", 3);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    const void *key = (const void *)(uintptr_t)args[1];
    void *out = (void *)(uintptr_t)args[2];
    if (!m || !key || !out || !spl_map_valid(m))
        return 1;
    uintptr_t i = spl_map_find_slot(m, key);
    if (i == ~(uintptr_t)0)
        return 1;
    memcpy(out, (void *)(m->vals.data + i * m->val_size), m->val_size);
    return 0;
}

static spl_vm_val_t vm_map_del(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_del", 2);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    const void *key = (const void *)(uintptr_t)args[1];
    if (!m || !key || !spl_map_valid(m))
        return 1;
    uintptr_t i = spl_map_find_slot(m, key);
    if (i == ~(uintptr_t)0)
        return 1;
    ((unsigned char *)m->states.data)[i] = 2;
    m->size--;
    return 0;
}

static spl_vm_val_t vm_map_clear(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_clear", 1);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    if (!m)
        return 0;
    if (m->states.data && m->states.len)
        memset((void *)m->states.data, 0, m->states.len);
    m->size = 0;
    return 0;
}

static spl_vm_val_t vm_map_drop(spl_vm_t *_vm, int nargs, spl_vm_val_t *args) {
    (void)_vm;
    CHECK_NARGS("vm_map_drop", 1);
    spl_map_t *m = (spl_map_t *)(uintptr_t)args[0];
    if (!m)
        return 0;
    spl_vec_release(&m->keys);
    spl_vec_release(&m->vals);
    spl_vec_release(&m->states);
    m->size = 0;
    return 0;
}

void spl_syscall_register(spl_prog_t *prog) {
    static spl_vm_native_t table[] = {
        /* OS operations */
        {"vm_exit", 0, vm_exit},
        {"vm_putchar", 0, vm_putchar},
        {"vm_getchar", 0, vm_getchar},
        {"vm_putint", 0, vm_putint},
        {"vm_putstr", 0, vm_putstr},
        {"vm_fopen", 0, vm_fopen},
        {"vm_fclose", 0, vm_fclose},
        {"vm_fread", 0, vm_fread},
        {"vm_fwrite", 0, vm_fwrite},
        {"vm_fsize", 0, vm_fsize},
        {"vm_stdin", 0, vm_stdin},
        {"vm_stdout", 0, vm_stdout},
        {"vm_stderr", 0, vm_stderr},
        {"vm_read_file", 0, vm_read_file},
        {"vm_alloc", 0, vm_alloc},
        {"vm_free", 0, vm_free},
        {"vm_realloc", 0, vm_realloc},
        {"vm_memcpy", 0, vm_memcpy},
        {"vm_memcmp", 0, vm_memcmp},
        {"vm_printf", 0, vm_printf},
        {"vm_vec_valid", 0, vm_vec_valid},
        {"vm_vec_push_raw", 0, vm_vec_push_raw},
        {"vm_vec_push", 0, vm_vec_push},
        {"vm_vec_get", 0, vm_vec_get},
        {"vm_vec_allocate", 0, vm_vec_allocate},
        {"vm_vec_reserve_raw", 0, vm_vec_reserve_raw},
        {"vm_vec_append", 0, vm_vec_append},
        {"vm_vec_set", 0, vm_vec_set},
        {"vm_vec_pop", 0, vm_vec_pop},
        {"vm_vec_clone", 0, vm_vec_clone},
        {"vm_vec_drop", 0, vm_vec_drop},
        {"vm_memmove", 0, vm_memmove},
        {"vm_bytes_eq", 0, vm_bytes_eq},
        {"vm_map_valid", 0, vm_map_valid},
        {"vm_map_reserve", 0, vm_map_reserve},
        {"vm_map_put", 0, vm_map_put},
        {"vm_map_get", 0, vm_map_get},
        {"vm_map_del", 0, vm_map_del},
        {"vm_map_clear", 0, vm_map_clear},
        {"vm_map_drop", 0, vm_map_drop},

        {"vm_timestamp", 0, vm_timestamp},
        /* VM operations (comptime) */
        {"vm_new", 0, vm_new},
        {"vm_drop", 0, vm_drop},
        {"vm_load", 0, vm_load},
        {"vm_load_from_buffer", 0, vm_load_from_buffer},
        {"vm_push", 0, vm_push},
        {"vm_call", 0, vm_call},
        {"vm_run", 0, vm_run},
    };
    int n = (int)(sizeof(table) / sizeof(table[0]));
    if (!prog)
        return;
    vec_for(prog->natives, i) {
        spl_vm_native_t *nat = &vec_at(prog->natives, i);
        if (!nat->name || nat->impl_fn)
            continue;
        for (int j = 0; j < n; j++) {
            if (strcmp(nat->name, table[j].name) == 0) {
                nat->impl_fn = table[j].impl_fn;
                break;
            }
        }
    }
}